Skip to main content
If you used Splox before, you built workflows: a canvas, nodes you dragged onto it, edges between them, an entry node at the top. None of that exists now. The product is an agent with its own Linux machine, and what the agent is, is Python in a git repository. This page is the map from the old vocabulary to the new one. Nothing here is a rename — most of these are a different mechanism that does the same job.

​
The vocabulary

The v1 endpoints behind the old canvas — /api/v1/workflows, /api/v1/memory/instances, /api/v1/mcp/connections — return 404. The current API is v2.

​
Why the canvas went

A graph draws the decisions in advance. The model makes them per message, with the whole conversation in front of it. Every branch that was drawn as a switch node was a decision taken away from the model and frozen at design time. What is left that a graph really did own — which agent takes this message — is now a function, because it is the one thing only your own code can know:
programs/splox/main.py
Code is reviewable and versioned. A harness is a git repository. A version is a commit. A run records the exact commit it executed, so you can read back the tree that produced any answer. A canvas had a version number; it did not have a diff. Tools compose when they are called in code. A JSON tool call gets one result back into the context window and the model pays for all of it. A tool that is a Python function can be chained with four others in one script, filtered, and made to print only the answer. That is why tools here are imports rather than calls. Hooks instead of forks. The turn loop has ten named points — context, model, tools, guard, memory, errors, stop — and a hook is a file that answers one of them. You change one decision without owning the loop, and a hook that fails falls back rather than taking the run down.

​
What has no replacement

There is no visual editor. If what you want is a fixed pipeline with phases and checkpoints, you write it: it is ordinary Python in a program, with while True, a queue, a retry and a database if it needs one. The starter tree carries programs/compiler/ as a worked example of that shape — several agents, phases, resumable — and its README is the thing to read before building anything similar.

​
If you are coming back

1

Your account already has a machine

Sign in and you have a harness copied from the Assistant starter and a machine running it. Quickstart.
2

Read the six words

Machine, harness, program, agent, run, chat. Concepts.
3

Port the one workflow you actually miss

Write it as a program. Patterns has four shapes to copy: a router, a bot, a nightly job and a fan-out.